Hunza Burushaski Research — ChatGPT Plugin Starter

Status: working starter repository, not deployed, not yet a public-ready plugin, no verified Burushaski linguistic content included. It provides a source-grounded research architecture, not a trained Burushaski language model.

What's included

• Portable plugin.json manifest and mcp.json Streamable HTTP endpoint configuration.

• skills/hunza-burushaski/SKILL.md for source-grounded research/translation workflow.

• Python Model Context Protocol server with six tools.

• SQLite schema with source rights, approvals, corrections, and audit votes.

• Offline CLI for adding sources, importing records, two-reviewer approval, corrections, and revocation.

• Research JSONL exporter with consent/rights filtering and source/speaker-group partitioning.

• Tests, Dockerfile, ethics/privacy notes, deployment guide, and data templates.

Related MCP server: srj-wikidata-mcp

Quick start (Python 3.11+)

cd burushaski-hunza-plugin
python -m venv .venv
# Activate: Linux/macOS: source .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
pip install -r requirements.txt
python -m burushaski.cli init
python -m burushaski.cli stats
python -m burushaski.server

The local server then listens on http://127.0.0.1:8000/mcp using Streamable HTTP. For tool discovery, connect an MCP inspector or compatible client; /mcp is an MCP endpoint, not a normal browser page. See docs/DEPLOYMENT.md.

Note: MCP uses Python SDK maintenance-line v1 (mcp>=1.28,<2). Porting to v2 requires updating imports and transport settings; see the official MCP SDK migration guide.

First real data, without fabricated translations

1. Identify a native Hunza Burushaski speaker and a separate pair of qualified reviewers.

2. Obtain and document permission for retrieval, training, and redistribution separately. The MIT code license does not grant permission to use a corpus.

3. Create a source record:

python -m burushaski.cli add-source \
  --id speaker-session-001 \
  --title "Consented elicitation session 001" \
  --creator "Pseudonym-of-consenting-speaker" \
  --license-id "Written project consent form v1" \
  --consent-reference "private-consent-register/session-001" \
  --can-retrieve
python -m burushaski.cli verify-source speaker-session-001 --rights-checked

4. Fill a copy of data/records_template.csv with genuinely verified content, using source_id=speaker-session-001, then import:

python -m burushaski.cli import-csv /path/to/your-private-records.csv

5. Each imported ID starts pending. Two different qualified reviewers must approve each one:

python -m burushaski.cli review-record RECORD_ID --reviewer reviewer-A --decision approve
python -m burushaski.cli review-record RECORD_ID --reviewer reviewer-B --decision approve

6. Search the data via MCP. Only records whose source is verified and allows retrieval, and that have two approvals, are visible.

Moderated corrections (offline workflow)

Create a JSON proposal file (example field names only, not real Burushaski):

{
  "kind": "lexicon",
  "burushaski_text": "REPLACE_WITH_SPEAKER_VERIFIED_FORM",
  "english_text": "REPLACE_WITH_GLOSS",
  "script": "roman"
}

Submit and approve it using trusted local access:

python -m burushaski.cli submit-correction --proposal-json /private/proposal.json \
  --contributor contributor-alias --consent-to-publish --rights-attested
python -m burushaski.cli review-correction CORRECTION_ID \
  --reviewer reviewer-A --decision approve --rights-checked
python -m burushaski.cli review-correction CORRECTION_ID \
  --reviewer reviewer-B --decision approve --rights-checked

To correct an existing approved record add --target ORIGINAL_RECORD_ID. Approved corrections create a new record version; the original becomes superseded. The submitted contributor alias must not be used as the reviewer identity. Training consent is always separate (--consent-to-training).

Future model training: data export, not training execution

Only source-verified, approved entries with can_train=1 and can_redistribute=1 enter the portable export. For example, --can-train --can-redistribute must be documented at source creation, or --consent-to-training provided for a community contribution; community contributions default to can_redistribute=0 and therefore need a separate license-clearing workflow before portable export.

python -m burushaski.cli export --output ./exports/research-v1

The exporter writes train.jsonl, validation.jsonl, test.jsonl, provenance.csv, and export_report.json. These are research data files, not a model fine-tuning job or a directly usable OpenAI fine-tuning upload. Group splits by speaker/source to reduce leakage; an empty evaluation split is a blocker for meaningful model evaluation.

Tests

python -m pip install pytest
python -m pytest -q

Before putting this online

• Replace https://replace-with-your-domain.example/mcp in mcp.json with your working HTTPS MCP endpoint.

• Use a durable disk/managed database and HTTPS reverse proxy. SQLite is for a single-instance pilot, not an autoscaled public system.

• Implement OAuth 2.1 user identification, scoped write access, abuse prevention and a privacy policy before accepting public corrections. Public submissions are disabled by default.

• Add correct publisher metadata, icon, hosted privacy/terms pages, domain verification, and review materials for public plugin submission.

• Confirm allowed uses of any externally published Burushaski corpora; do not assume academic access permits redistribution or commercial training.

See docs/TECHNICAL_BLUEPRINT.md, docs/DEPLOYMENT.md, docs/RESEARCH_AND_EVAL.md, and docs/SAFETY_PRIVACY.md.

Sources for the integration (verified November 2026)

• OpenAI plugin packaging: https://developers.openai.com/plugins/build/plugins

• MCP server guide: https://developers.openai.com/plugins/build/mcp-server

• ChatGPT plugin quickstart: https://developers.openai.com/plugins/quickstart

• Plugin submission: https://developers.openai.com/plugins/deploy/submission

• MCP SDK v1 reference: https://py.sdk.modelcontextprotocol.io/v1/

• Model training availability: https://developers.openai.com/api/docs/guides/supervised-fine-tuning